distillation-and-post-training
MiniLLM, DPKD, and Tuna distillation and ranking-finetuning workflows for LMOps.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- distillation-and-post-training
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- MiniLLM, DPKD, and Tuna distillation and ranking-finetuning workflows for LMOps.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/dpkd-workflows.md4.5 KB
- 参考文档references/minillm-workflows.md4.9 KB
- 参考文档references/troubleshooting.md4.4 KB
- 参考文档references/tuna-data-and-training.md4.6 KB
- 脚本scripts/check_tuna_ranking_data.py11.7 KB
- 脚本scripts/model_parallel_conversion_plan.py10.3 KB
03 SKILL.md 正文
Distillation and Post-Training
Use this sub-skill for MiniLLM, DPKD, and Tuna workflows that distill or rank-finetune language models without entering VeRL-style experiential learning.
Route here when
- The task is MiniLLM SFT, KD, SeqKD, on-policy MiniLLM training, evaluation, exposure-bias analysis, data processing, model resource planning, or tensor parallel conversion.
- The task is DPKD training or evaluation, including the runner scripts, argument-group choices, DPO-style switches, and checkpoint/output planning.
- The task is Tuna probabilistic ranking or contextual ranking data preparation, GPT-4 provenance checks, or ranking-aware finetuning.
Route away
- GAD, OEL, OPCD, LLM-as-a-Coach, and OPO belong in
../rl-experiential-learning/SKILL.md. - ResLoRA belongs in
../adaptation-and-training/SKILL.md.
Bundled helpers
scripts/model_parallel_conversion_plan.pypreflights checkpoint family, source/target MP sizes, and path layout before any conversion.scripts/check_tuna_ranking_data.pyvalidates the minimal probabilistic and contextual ranking JSON shapes used by Tuna.
Read next
references/minillm-workflows.mdreferences/dpkd-workflows.mdreferences/tuna-data-and-training.mdreferences/troubleshooting.md
Working order
- Decide whether the request is MiniLLM, DPKD, or Tuna.
- Check the family reference and the bundled helper before any path-sensitive or schema-sensitive action.
- Treat large model downloads, multi-node launches, and OpenAI-backed ranking generation as documented workflow steps, not cheap smoke tests.
- If a failure looks like a shared LMOps issue, consult the parent skill's
broad project index when available at
../../references/project-index.md.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).